© 2018 - Analytics Link

A huge leap for A.I. as AlphaGo Zero wipes the floor with AlphaGo 100-0

October 19, 2017

Cast your mind back to March 2016.  You probably remember hearing about AlphaGo, the A.I system developed by the Google owned DeepMind that beat 18 time world champion Go Grandmaster Lee Sedol 4-1 in a $1 million best of five match.  

 

This was an amazing feat for the world of A.I and Deep Learning, how could we possibly progress further in a game that has more legal positions than there are atoms in the universe..?

 

 - Lee Sedol looking slightly disheartened during his 4-1 loss to the original AlphaGo in March 2016

 

 

Well scratch that - AlphaGo has recently lost 100-0 to an upgraded version of itself...Introducing AlphaGo Zero.

 

All previous versions of AlphaGo started by training from human data, i.e. they were told "from this position, the human expert played this move, and it this position, the human expert used this particular move"

 

AlphaGo Zero doesn't use any human data and therefore is no longer constrained by the limits of human knowledge, it instead has to learn from itself completely through self-play.  Through millions of iterations it has accumulated thousands of years of human game knowledge, sometimes discarding complex human moves for unconventional strategies and self developed, creative new moves.

 

A snippet from the DeepMind release states that “It is able to learn tabula rasa [from a clean slate] from the strongest player in the world: AlphaGo itself.”

 

AlphaGo Zero's learning curve - source: DeepMind

 

To learn more, I'd recommend reading through the original source paper here, which includes some great explanations and visualisations

 

Also, on YouTube you can find Professor David Silver, the lead researcher for AlphaGo lends his thoughts in the below videos - well worth a watch!

 

Starting from scratch

 

"People tend to assume that Machine Learning is about Big Data and massive amounts of computation. Actually what we saw in AlphaGo Zero is that algorithms matter much more than either compute or data availability" - Professor David Silver

 

Discovering new knowledge

 

 

 

 

 

 

Andrew Jones has over 10 years experience in Analytics, Data Science, and Machine Learning within the Digital, Retail, Telecoms and Travel industries - most recently at Amazon, & Sony PlayStation

 

 

 

 

 

We hope you enjoyed the read.  Please do share with the social buttons below and let us know any thoughts on LinkedIn!

 

 

Share on Facebook
Please reload

Please reload

Recent Posts